Obstacle  Detection  through  Forward  MovementQiuwei Shou and  Alan  Wu,  CSCI-­‐B657  Computer  Vision,  Indiana  University

• Obstacle   detection   and  avoidance   are   essential  capabilities  for   autonomous  robots  (land  and  air)

• While  traditional   multi-­‐sensory   approach   addresses  obstacle   detection,   the   use  of  a  single  (monocular)  camera   enables   a  more   affordable   solution

• Computational  requirements   must  be  manageable   by  embedded   controllers

• Project   is  inspired  by  [Mori  and  Scherer,   2013]1 paper

• We   manually  assembled  our   image  dataset,   which  is  primarily   derived   from  drone  hobbyist  videos  taken   from  a  first-­‐person   view   (FPV),  downloaded  from   YouTube

• Mostly  forest   scenes   with  various   levels  of  luminance  and  tilt  angles   throughout   flight

• Two  datasets   totaling   ~500  previous/current   frame  combinations

• Set   1:  ~400  dense  forest• Set   2:  ~100  open  field  (no  obstacles)

Feature   Detection

Feature   Filtering

Template   Extraction Feature   Matching

Obstacle   Detection   Algorithm

Previous   Image

Obstacle:   Yes   or  No

Current  Image

• We  divide  each  image  into  4x4  regions  and  use  SURF  for  fast  feature  detection• Only  certain  features  in  the  “previous”  frame  are  selected  for  matching  with  

“current”  frame• Features   must  be  of  a  minimum   size• Features   must  be  within   certain   regions   of   the   image

• We  then  perform  template  matching1 to  detect  approaching  object(s)  that  may  be  obstacles.    See  upper-­‐right  of  poster  for  details• Initial   template  window   size   W  affects  detection   accuracy  as   seen   in  graph  below

Motivation

Dataset

Methodology

Current  ImagePrevious  Image

Sample   Obstacle  Dataset

Sample   Non-­‐obstacle  Dataset

FeatureFiltering

TemplateExtraction

Challenges• Dataset   is   largely   “uncalibrated”   – frame  captures   estimated   to  

contain  obstacles,   but   not  guaranteed

• SURF   is  not  optimal   with   noisy   background    

• Often  unable   to  differentiate   features   on   the  ground   from  protruding   objects   (such  as   trees),   especially   when   the  camera   is  tilted   downward

References:  1Mori,   Tomoyuki and   Sebastian   Scherer.    “First   Results  in  Detecting   and   Avoiding  Frontal  Obstacles   from  a  Monocular  Camera   for  Micro  Unmanned   Aerial   Vehicles.”  ICRA,   2013.

2http://www.ingeniousingenuity.com3http://clipartion.com/

2

3

Conclusion  and  Future  Work• Able   to  detect  oncoming   obstacles   using   a  low   computation  

algorithm   based   on  SURF   features  and   template  matching

• Access   textural   library   (ex.  tree  bark)   for  more   targeted  detection  and  matching;  possibly   combine   with   Visual   SLAM   to  achieve  higher  fidelity   in  estimating   obstacle   depth

Template  Matching• Perform  template  matching1 to  determine  scale  increase  of  

approaching  object• For  each   remaining   feature   after   f iltering   in  “previous”   frame,   a   small  

window  is  drawn  around  feature• A  similar   window  is  drawn  around  matching  feature   in  “current”  frame,   with  

window  size  scaled   up  by  K;  resize   window  in  “previous”  frame   also   by  K• Scalar   K  is  determined   such  that  the  correlation   distance   (SSD)   between  

“previous”   and  “current”   windows  is  minimized.    Discard   any   matches   with  distance   D   >  Dmax and  K  <   Kmin

Previous   Frame  Feature Current   Frame  Feature

1.  Scale   window  size   by  K

2.  Resize   feature  window  by  K

3.  Calculate   distance   for  various  K

Results• We  achieved  an  accuracy  of  84% (i.e.  False  Neg =  16%)   in  detecting  

obstacles   in  Set  1   (dense   forest)   and  a  False   Pos rate  of  33% in  falsely   detecting  obstacles   in  Set  2   (open   field),   for   initial   window  size   of  30  and   ¼  resolution   relative   to  original   (72  ppi)

• Plots   show   effects  of  initial window   size   W  and   image   resolution

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

10 15 20 25 30 35 40 45 50

False

-Rate

Template-Window-Size

Template-Window-Size-vs.-Detection-Accuracy-on-Baseline-Dataset

False-Neg

False-Pos

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

20 25 30 35 40

False/Ra

te

Template/Window/Size

Template/Window/Size/vs./Detection/Accuracy,/0.25x/ppi/

False/Neg

False/Pos

0

0.1

0.2

0.3

0.4

0.5

0.6

0 0.1 0.2 0.3 0.4 0.5

False

.Rate

Resolution. re:.Original

Resolution.vs..Detection.Accuracy,.W.=.30

False.Neg

False.Pos